Commonsense Go

[djhbrown]

https://youtu.be/Nns70047Rxo

In this video, design of software able to think and talk about Go in a
commonsense way,is described and illustrated by two examples:
1. solving a tactical problem presented by Jennie Shen of Guo Juan's
Internet go school.
2. figuring out how to respond to Lee Sedol's magic wedge in game 4 of
his match with
Alphago.

Thank you to Guo Juan for permission to reproduce Jennie Shen's
excellent lecture.
And to Walter Reitman and Bruce Wilcox for their seminal work and moral support.

[Krama]

https://youtu.be/Nns70047Rxo

In this video, design of software able to think and talk about Go in a
commonsense way,is described and illustrated by two examples:
1. solving a tactical problem presented by Jennie Shen of Guo Juan's
Internet go school.
2. figuring out how to respond to Lee Sedol's magic wedge in game 4 of
his match with
Alphago.

Thank you to Guo Juan for permission to reproduce Jennie Shen's
excellent lecture.
And to Walter Reitman and Bruce Wilcox for their seminal work and moral support.

Where can we download the software?

[hyperpape]

Where can we download the software?

What a fascinating question

[uPWarrior]

TLDR: A 30 minutes video of computer go ideas from the 1990's.

[djhbrown]

Monte Carlo and DCNN leave the Go programs of the 1990s in the dust.

The basic principles of AI planning and goal-directed problem solving have been around since the 1970s, but, with the exception of Wilcox's sector lines, none of the ideas in my video have ever been implemented in any previous Go program, nor discussed in any academic paper other than my own.

The design of Commonsense Go includes new algorithms for computing territory and influence, circumscribing groups and assessing their strengths, and new deductive methods for strategic and tactical reasoning.

Unlike all other Go programs to date, it can talk about what it's thinking in plain English, so would thus be useful to people learning Go.

Alphago has the fundamental flaw that it doesn't "know" what it's doing, because neither DCNN nor Monte Carlo embodies basic commonsense intelligence of being able to construct and reason about a meaningful computational model of the dynamics of the world.

Commonsense Go does "know". It can can find a better move sequence (see board image) than Alphago did at move 79 in game 4 of its match with Lee Sedol.



The program is designed, but it's up to others to turn it into software. Could it beat Alphago across the board, and with much less hardware? I think it could.

pdf version of the video, designed to be read in "dual mode" (ie, 2 pages at a time) with odd-numbered pages on the left:

http://papers.ssrn.com/sol3/papers.cfm? ... id=2818149

[wineandgolover]

The program is designed, but it's up to others to turn it into software. Could it beat Alphago across the board, and with much less hardware? I think it could.

Then it would be worth a fortune. Self fund the development and get rich! Maybe Google would buy your company!

[Uberdude]

The program is designed, but it's up to others to turn it into software. Could it beat Alphago across the board, and with much less hardware? I think it could.

I think it wouldn't as you'd be making a bot similar to GnuGo. Trying to build strong bots that embody human-like reasoning about meaningful concepts has not been successful for Go. You might be able to make a useful teaching tool for weak players, but not a strong player.

[pookpooi]

djhbrown you might be interest in reading this https://arxiv.org/pdf/1604.04315v2

[RobertJasiek]

The design of Commonsense Go includes new algorithms for computing territory and influence, circumscribing groups and assessing their strengths, and new deductive methods for strategic and tactical reasoning.

No, because what is described (in the paper, I have not viewed the video) is not influence and strength. Concepts should also be based on connection status instead of proximity sold as "influence".

[djhbrown]

you'd be making a bot similar to GnuGo

no, you wouldn't. i do call it a proposal for a new Gnu Go, because i support the principles of the Gnu Foundation. But beyond that, there is no similarity between the programs whatsoever that is not also shared by Alphago and every other bot ever written to play Go. Not a single jot. If anyone can find one, do tell.

https://arxiv.org/pdf/1604.04315v2

i agree with their observations about Alphago and Watson. Commonsense Go (CG) does not embrace all human general commonsense, which is much richer and broader; it only embraces a version of the sense about Go that is common among Go players. But it is a somewhat stylised version of that sense; for example, it includes the concepts of clusters, colour maps and shadow maps, all of which are novel. I do not know whether a pro would see what CG calls a group in quite the same way. CG would not be a candidate for the Allen AI Challenge as it cannot answer arbitrary questions. However, it does "move beyond information retrieval and into intelligent inferencing over multiple facts, concepts, and relationships to produce correct, explainable answers to these questions" - subject to the constraint that the only questions it can answer are "Why?" and "What do you think if black/white plays <here> now?".

what is described (in the paper, I have not viewed the video) is not influence and strength. Concepts should also be based on connection status instead of proximity sold as "influence"

influence is computed as g(shadow map). group strength has one of 4 qualitative values: alive, dead, strong and weak. Colour and shadow concepts ARE based on potential connectivity, not merely proximity, as you will see if you study the algorithms more carefully.

It should be obvious that CG's perceptions of territory and influence are its perceptions, and only its perceptions. No claim is made that CG's perceptions are mathematical proofs of unconditional connectedness, but i do claim that they correlate well with what a pro would say. If you know a pro, you can ask them. I have asked 3 pros what they think of CG's reply to Lee Sedol's move 78, but have yet to receive an answer. CG can be said to "believe" that clusters are tightly connected, and groups more loosely so. If you watch the video, you will see examples of how the colour and shadow maps are iteratively computed and recomputed after dead clusters are identified. The perimeters of clusters and groups are not impenetrable walls as they are preceived without dynamic analysis. However, the metamethod does perform dynamic analysis of connection status when applying the methods "connect" and "separate".

Robert, i believe you have some ideas of your own about how to compute influence. Please apply them to the examples in the paper/video, and share the results with us, so we can see how what your ideas produce compares with what CG thinks.

[Uberdude]

you'd be making a bot similar to GnuGo

no, you wouldn't. i do call it a proposal for a new Gnu Go, because i support the principles of the Gnu Foundation. But beyond that, there is no similarity between the programs whatsoever that is not also shared by Alphago and every other bot ever written to play Go. Not a single jot. If anyone can find one, do tell.

Approaching the game in terms of concepts like groups.

[Javaness2]

troll troll troll your boat gently down the stream

[RobertJasiek]

djhbrown, influence concepts would be an overkill to the paper's positions. It should be sufficient to assess connection and life statuses, Current Territory and updates of these after possible next moves. (I lack time for studying these in those positions. If you want to know, apply my terms to them.)

[pookpooi]

troll troll troll your boat gently down the stream

hmmm I don't think this is a troll thread, after all, we already 'had' someone who claim that his program can beat AlphaGo, and the admin quickly ban him, that is the real troll.

[oren]

troll troll troll your boat gently down the stream

Yeah, I kind of vote for not a troll... just someone who doesn't understand computer go as well as he thinks he does.